Complex-Valued Neural Network And Its Application In Speaker Recognition

Complex-valued neural networks are put forward in contrast to real-valued neural networks. The traditional real-valued neural networks, whose inputs, network parameters (such as weights and thresholds), outputs are all real-valued. However in many practical applications input and output data are complex, such as communications, electromagnetics, image processing, speech processing, and other fields. The approach of traditional neural networks to process complex data is extracting the real part and the imaginary part to be addressed. That is, a complex data is separated into two real data, and then use two real-valued neural networks to be addressed. Complex-valued neural networks, whose inputs, parameters and outputs are all complex valued, do not need to separate the real part and the imaginary part to deal with, instead, as a complex whole data to deal with, so it is more natural and effective. For processing signals such as voice signals and the like on the "wave", the complex neural network has a great advantage. This thesis researches on complex-valued neural networks and its application in speaker recognition. Main works focused on the following three aspects.Firstly, this thesis analyzes the basic structure of complex-valued neural networks, and studies the commonly used functions of complex-valued neurons. It studies the learn rules and algorithms of complex-valued BP neural networks, based on the real-valued BP neural networks. The thesis also improves complex-valued BP neural network algorithm based on the analysis of the defects in complex-valued BP neural network algorithm. This thesis has experimental simulation, and the results show that complex-valued networks have better performance compared with the real-valued BP neural networks.Secondly, this thesis designs the complex-valued coefficient FIR digital filter with complex-valued neural networks, based on the analysis of the advantages of complex-valued BP neural networks. Neural networks have a strong nonlinear mapping ability, so they can design the complex-valued coefficient FIR digital filter of any frequency response. In this thesis, examples of optimized design of the FIR low-pass filter and band-pass filter based on the complex-valued neural networks are also presented. Simulation results show that, comparing with real-valued neural networks, the complex-valued neural network is more natural and effective for the designing of FIR filter, and has higher training speed and stronger generalization ability.Finally, according to the speaker's voice characteristics, this thesis designs the speaker recognition system and its algorithm with complex-valued neural networks. This thesis uses the complex-valued BP neural networks as classification, recognition model. The most commonly used speech feature parameters are mel-frequency cepstral coefficients (MFCC) and linear prediction cepstral coefficients (LPCC). But these two characteristic parameters of speech have been ignored the phase information of speech signal during extracting the speech cepstrum. Taking into account complex-valued neural networks'advantages to process the complex data, this thesis uses the complex cepstrum of speech signals (it retains the phase information of the original speech signal) as a characteristic parameter to identify and validate.